Machine learning-based airborne laser radar forest stock estimation method and device

Abstract

The invention discloses an airborne laser radar forest stock estimation scheme based on machine learning, and the scheme comprises the steps: generating a structured modeling data set according to various types of sample plot point cloud data collected by an airborne laser radar and sample plot indexes obtained through the actual investigation of various sample plots, carrying out the division of the data set through employing various random seeds, and carrying out the calculation of the divided data set; a plurality of modeling data set division results are obtained; performing grid search on the machine learning model parameters to obtain a plurality of model parameter combinations; training a machine learning model corresponding to the model parameter combination by adopting a K-fold cross validation method and a plurality of modeling data set division results to obtain a trained machine learning model, and testing the trained machine learning model through a test data set to obtain an evaluation index; and determining a target machine learning model based on the evaluation index corresponding to each model parameter combination, training the target machine learning model based on the structured modeling data set, and generating a stock estimation model, the stock estimation model being suitable for various forest types, and the predicted stock result being accurate.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of forestry informatization and remote sensing resource monitoring, and particularly relates to a method and device for estimating forest stock volume based on machine learning. BACKGROUND

[0002] Forest stock volume is one of the core indicators of forest resource quantity and ecological service capacity, and can reflect the total scale and level of forest resources in a country or region, and plays an important supporting role in many fields such as national forest resource monitoring, carbon sink accounting, forest management and ecological quality assessment. The traditional forest stock volume estimation method mainly relies on ground plot investigation and regression model. Although it has certain accuracy in a small range, it has problems such as long investigation period, high labor intensity, and insufficient representativeness of sample plots, and it is difficult to meet the actual needs of high efficiency, high precision and wide coverage estimation technology for current forest resource management. In addition, the traditional modeling method generally relies on the idea of classifying modeling according to tree species or forest types, and multiple sub-models are constructed between different types of sample plots, which not only has a complicated modeling process, but also lacks uniformity and universality among models, resulting in weak ability of promotion and application at regional scale, and difficulty in adapting to the needs of rapid assessment of large-scale and multi-type forest land.

[0003] With the development of remote sensing technology, especially laser radar technology, airborne laser radar as an active three-dimensional remote sensing means can penetrate the forest canopy and obtain vertical structure information, providing a new data basis for forest resource estimation. The stock volume modeling method based on laser radar has become a hot spot in current forestry remote sensing research and application. However, most of the current researches still use traditional statistical regression method for modeling, and the model structure is simple, which is difficult to fully excavate the non-linear relationship between high-dimensional point cloud variables and stock volume; at the same time, the model has poor adaptability to forest type difference, terrain change and other factors, resulting in unstable estimation accuracy and insufficient generalization ability.

[0004] Therefore, the existing traditional forest stock volume estimation method and the stock volume estimation method based on laser radar cannot obtain high-accuracy forest stock volume estimation results. SUMMARY

[0005] The purpose of the embodiment of the present application is to provide a method and device for estimating forest stock volume based on machine learning of airborne laser radar, and an electronic equipment, which can solve the problem of low accuracy of forest stock volume estimation results in the prior art.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The embodiment of the present application provides a method for estimating forest stock volume based on machine learning of airborne laser radar, wherein the method comprises:

[0008] According to the point cloud data of various types of sample sites collected by the airborne laser radar and the sample site indexes obtained by the actual investigation of the sample sites, a structured modeling data set is generated;

[0009] The structured modeling data set is divided by using multiple random seeds to obtain multiple modeling data set division results, and each modeling data set division result includes a first preset proportion of training data set and a second preset proportion of test data set;

[0010] A plurality of preset machine learning models are determined, wherein the machine learning models include a random forest regression model, a gradient boosting tree model, and a classification and regression tree model;

[0011] For each machine learning model, a grid search is performed on the machine learning model parameters to obtain multiple model parameter combinations;

[0012] For each model parameter combination, a K-fold cross-validation method and the multiple modeling data set division results are used to train the machine learning model corresponding to the model parameter combination to obtain a trained machine learning model; the trained machine learning model is tested by using the test data set to obtain an evaluation index;

[0013] Based on the evaluation index corresponding to each model parameter combination, a target machine learning model is determined;

[0014] Based on the structured modeling data set, the target machine learning model is trained to generate an accumulation amount estimation model.

[0015] Optionally, the step of generating the structured modeling data set based on the point cloud data of various types of sample sites collected by the airborne laser radar and the sample site indexes obtained by the actual investigation of the sample sites includes:

[0016] For each sample site, after the point cloud data collected by the airborne laser radar is spliced and geometrically corrected based on the flight control points and the ground control points, it is converted to a specific projection coordinate system;

[0017] The flight height, scanning angle, and ground object reflectivity are used to correct the laser intensity;

[0018] According to the set height and reflectivity threshold, abnormal point cloud data is removed to obtain preprocessed high-dimensional point cloud data;

[0019] From the high-dimensional point cloud data, a preset structural feature variable is extracted to obtain high-precision remote sensing point cloud data.

[0020] Optionally, the step of dividing the structured modeling data set by using multiple random seeds to obtain multiple modeling data set division results includes:

[0021] A plurality of different stratified sampling strategies are adopted, and the structured modeling data set is stratified sampled according to preset key variables, so as to obtain a plurality of modeling data set division results.

[0022] Optionally, for each machine learning model, the grid search is performed on the machine learning model parameters, so as to obtain a plurality of model parameter combinations.

[0023] Receiving a selection operation of a user on parameters;

[0024] Determining whether the number of parameters selected by the user exceeds a preset number;

[0025] If the number of parameters selected by the user does not exceed the preset number, the parameters selected by the user are arranged and combined, so as to obtain a plurality of model parameter combinations.

[0026] If the number of parameters selected by the user exceeds the preset number, a preset number of parameters are randomly selected from the parameters selected by the user, and the parameters are arranged and combined, so as to obtain a plurality of model parameter combinations.

[0027] Optionally, for each model parameter combination, the K-fold cross-validation method and the plurality of modeling data set division results are adopted to train the machine learning model corresponding to the model parameter combination, so as to obtain a trained machine learning model, and the step includes:

[0028] Determining a K value in the K-fold cross-validation method;

[0029] Dividing a training data set into K equal-size subsets;

[0030] In each round of training of the machine model corresponding to the model parameter combination, one of the subsets is used as a test data set, and the remaining K-1 subsets are used as training data sets for single-round training;

[0031] The machine model corresponding to the model parameter combination is trained for K times, so as to obtain a trained machine learning model.

[0032] Optionally, based on the evaluation indexes corresponding to each model parameter combination, the step of determining a target machine learning model includes:

[0033] For each machine learning model, based on the evaluation indexes corresponding to each model parameter combination, the optimal model parameter combination corresponding to the machine learning model is determined, and a first target machine learning model is generated;

[0034] According to the evaluation indexes, an evaluation index optimal model is selected from the first target machine learning models as a target machine learning model.

[0035] The embodiment of the application also provides an airborne laser radar forest volume estimation device based on machine learning.

[0036] a generating module configured to generate a structured modeling dataset according to each type of sample plot cloud data collected by the airborne laser radar and sample plot indexes obtained by actual investigation of each sample plot;

[0037] a dividing module configured to divide the structured modeling dataset by using a plurality of random seeds to obtain a plurality of modeling dataset division results, each modeling dataset division result containing a first preset proportion of training dataset and a second preset proportion of testing dataset;

[0038] a determining module configured to determine a plurality of preset machine learning models, wherein the machine learning models include a random forest regression model, a gradient boosting tree model and a classification and regression tree model;

[0039] a searching module configured to perform grid search on parameters of each machine learning model to obtain a plurality of model parameter combinations;

[0040] a cross-validation module configured to train, for each model parameter combination, a machine learning model corresponding to the model parameter combination by using a K-fold cross-validation method and the plurality of modeling dataset division results to obtain a trained machine learning model, and test the trained machine learning model by using the testing dataset to obtain an evaluation index;

[0041] a screening module configured to determine a target machine learning model based on the evaluation index corresponding to each model parameter combination;

[0042] a training module configured to train the target machine learning model based on the structured modeling dataset to generate a storage capacity estimation model.

[0043] Optionally, the generating module includes:

[0044] a first sub-module configured to, for each sample plot, convert, after strip splicing and geometric correction of point cloud data collected by the airborne laser radar based on a flight control point and a ground control point, the point cloud data to a specific projection coordinate system;

[0045] a second sub-module configured to correct laser intensity by using flight height, scanning angle and ground object reflectivity;

[0046] a third sub-module configured to remove abnormal point cloud data according to a set height and reflectivity threshold to obtain preprocessed high-dimensional point cloud data;

[0047] a fourth sub-module configured to extract a preset structural feature variable from the high-dimensional point cloud data to obtain high-precision remote sensing point cloud data.

[0048] Optionally, the dividing module is specifically configured to:

[0049] A plurality of modeling data set division results are obtained by using a plurality of different stratified sampling strategies and stratified sampling of the structured modeling data set according to preset key variables.

[0050] Optionally, the search module comprises:

[0051] The fifth sub-module is configured to receive a selection operation of the user on the parameters.

[0052] The sixth sub-module is configured to determine whether the number of the parameters selected by the user exceeds a preset number.

[0053] The seventh sub-module is configured to, if the number of the parameters selected by the user does not exceed the preset number, perform permutation and combination on the parameters selected by the user to obtain a plurality of model parameter combinations.

[0054] The eighth sub-module is configured to, if the number of the parameters selected by the user exceeds the preset number, perform permutation and combination on a preset number of parameters randomly selected from the parameters selected by the user to obtain a plurality of model parameter combinations.

[0055] Optionally, the cross-validation module is specifically configured to:

[0056] Determine a K value in a K-fold cross-validation method.

[0057] Divide a training data set into K equal-size subsets.

[0058] In each round of training of the machine model corresponding to the model parameter combination, use one of the subsets as a test data set, and use the remaining K-1 subsets as training data sets for single-round training.

[0059] Repeat the training of the machine model corresponding to the model parameter combination for K times to obtain a trained machine learning model.

[0060] Optionally, the screening module is specifically configured to:

[0061] For each machine learning model, determine an optimal model parameter combination corresponding to the machine learning model based on evaluation indexes corresponding to various model parameter combinations to generate a first target machine learning model.

[0062] Select an evaluation index optimal model from the first target machine learning models as a target machine learning model according to the evaluation indexes.

[0063] An electronic device is provided, which comprises a processor, a memory, and a program or instruction stored on the memory and executable on the processor, and the program or instruction is executed by the processor to implement the steps of any one of the above-described machine learning-based airborne laser radar forest volume estimation methods.

[0064] The embodiment of the present application provides a readable storage medium, the readable storage medium stores programs or instructions, and the programs or instructions are executed by a processor to realize the steps of any one of the machine learning-based airborne laser radar forest volume estimation methods.

[0065] The machine learning-based airborne laser radar forest volume estimation scheme provided in the present application generates a structured modeling data set according to various types of sample point cloud data collected by an airborne laser radar and sample site indicators obtained by actually investigating various sample sites; a plurality of random seeds are used to divide the structured modeling data set, to obtain a plurality of modeling data set division results; a plurality of preset machine learning models are determined; for each machine learning model, grid search is performed on machine learning model parameters, to obtain a plurality of model parameter combinations; for each model parameter combination, a K-fold cross-validation method and the plurality of modeling data set division results are used to train a machine learning model corresponding to the model parameter combination, to obtain a trained machine learning model; the trained machine learning model is tested by using a test data set, to obtain evaluation indexes; a target machine learning model is determined based on the evaluation indexes corresponding to each model parameter combination; and the target machine learning model is trained based on the structured modeling data set, to generate a volume estimation model. The scheme provided in the present application has the following advantages. First, the scheme supports modeling without removing abnormal point cloud data, fully utilizes original data, can guarantee that the volume estimation model still maintains good performance in the case of containing extreme samples, and can improve the robustness and application range of the prediction result of the volume estimation model. Second, the scheme can be applied to various forest types without classifying modeling of various types of sample sites, and has good universality. Third, the scheme uses a plurality of random seeds and performs grid parameter search by using a plurality of parameter combinations, to further enhance the reliability and stability of the volume estimation model. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 is a step flowchart of a machine learning-based airborne laser radar forest volume estimation method according to an embodiment of the present application;

[0067] Figure 2 is a process schematic diagram of a machine learning-based airborne laser radar forest volume estimation method according to an embodiment of the present application;

[0068] Figure 3 is a structural block diagram of a machine learning-based airborne laser radar forest volume estimation device according to an embodiment of the present application. DETAILED DESCRIPTION

[0069] To make the technical problems, technical solutions and advantages of the present application clearer, the following will be described in detail with reference to the drawings and specific embodiments.

[0070] In recent years, with the development of artificial intelligence and data science, machine learning, as a data-driven modeling technique, has been widely applied in remote sensing data mining, feature recognition, and parameter inversion. Compared with traditional regression models, machine learning algorithms have shown significant advantages in handling high-dimensional, multi-variable, and complex non-linear data sets, enabling more comprehensive mining of potential patterns in the data and improving model prediction performance and stability. In forestry remote sensing applications, laser radar point cloud data has characteristics such as high dimension, complex structure, and variable redundancy. Using machine learning methods can automatically complete feature selection, parameter optimization, and error control, improving the accuracy and efficiency of forest structure parameter estimation. Therefore, by using advanced machine learning algorithms and integrating high-dimensional structure information from laser radar, this application proposes a volume estimation method with strong generalization ability and high modeling accuracy to break through the bottleneck of existing technology in terms of precision, efficiency, and adaptability.

[0071] Laser radar data has a high dimension and contains a large number of height and intensity statistical variables. However, current variable selection often relies on human experience or simple correlation analysis, lacking a systematic variable importance identification mechanism, which can easily miss key variables or introduce redundant information, affecting model performance and reducing estimation stability and efficiency. During model training, existing research often lacks scientific parameter optimization strategies, and the modeling method and process lack automation and standardization, which can easily cause overfitting or unreasonable parameters, affecting the generalization ability and reproducibility of the model on different data sets. The embodiments of the present application provide a volume estimation scheme with strong generalization ability and high modeling accuracy by using advanced machine learning algorithms and integrating high-dimensional structure information from laser radar, to break through the bottleneck of existing technology in terms of precision, efficiency, and adaptability.

[0072] The embodiments of the present application provide a machine learning-based airborne laser radar forest volume estimation scheme, which will be described in detail below in conjunction with the drawings, specific embodiments, and application scenarios.

[0073] As shown in the accompanying Figure 1 The machine learning-based airborne laser radar forest volume estimation method of the embodiments of the present application includes the following steps:

[0074] Step 101: Generate a structured modeling data set based on the various types of sample point cloud data collected by the airborne laser radar and the sample site indicators obtained from the actual investigation of each sample site.

[0075] The machine learning-based airborne laser radar forest volume estimation method provided by the embodiments of the present application can be applied to an electronic device, which includes a processor and a storage medium. The storage medium stores a computer program related to machine learning-based airborne laser radar forest volume estimation. The processor of the electronic device executes the computer program to implement the machine learning-based airborne laser radar forest volume estimation process.

[0076] In an alternative embodiment, the structured modeling data is generated according to the various types of sample plot cloud data collected by the airborne laser radar and the sample plot indicators obtained by the actual investigation of the various sample plots, and the generation manner comprises the following sub-steps:

[0077] Sub-step 1: For each sample plot, the point cloud data collected by the airborne laser radar is converted to a specific projection coordinate system after the flight control point and the ground control point are used to perform strip splicing and geometric correction.

[0078] The specific projection coordinate system can be the CGCS2000 projection coordinate system.

[0079] Sub-step 2: The flight height, scanning angle and ground object reflectivity are used to correct the laser intensity.

[0080] Sub-step 3: According to the set height and reflectivity threshold, the abnormal point cloud data is removed to obtain the preprocessed high-dimensional point cloud data.

[0081] Sub-step 4: The preset structural feature variables are extracted from the high-dimensional point cloud data to obtain high-precision remote sensing point cloud data.

[0082] The representative structural feature variables are extracted from the high-dimensional point cloud data, and the regression model can be constructed flexibly by selecting variables, which is not constrained by the number of variables.

[0083] An example of extracting the preset structural feature variables can be: 98 point cloud structural feature variables are extracted, including height type indicators (point cloud height average value, square average value, cubic average value, point cloud height skewness, point cloud cumulative height, etc.), intensity type indicators (point cloud intensity average value, median, point cloud intensity 30%, 40%, 50%, 60%, 70%, 75%, 80% quantile, etc.).

[0084] In the original point cloud data obtained in the data acquisition stage, there can be systematic errors, coordinate drift and intensity drift and other problems. The original point cloud data is preprocessed by the alternative embodiment to generate high-precision remote sensing point cloud data, which can ensure the data quality of the point cloud data and the accuracy of subsequent variable extraction.

[0085] In the sample plot data processing link in the embodiments of the present application, a strategy of retaining all data points, including extreme values and abnormal samples, is adopted, and a unified model system is constructed by fully utilizing the full amount of data, which is specifically embodied in:

[0086] Data fusion: The high-precision remote sensing point cloud data is accurately spatially matched and fused with the plot indicators obtained through field investigation to form a structured modeling data set, ensuring the integrity and consistency of the input data. The modeling process does not rely on classification information such as tree species, forest type or site type, but is based on the overall forest resource status to build the model. This makes the model more widely applicable and able to cover a variety of forest types and complex stand structures. During the data aggregation stage, all extreme values and abnormal samples are retained rather than being removed or filtered. This enhances the robustness of the model, allowing it to maintain good prediction performance when faced with complex and heterogeneous stand structures, improving the model's adaptability to forest resource diversity and providing a solid data foundation and technical support for accurate evaluation of different types of forest resources.

[0087] Step 102: Divide the structured modeling data set using multiple random seeds to obtain multiple modeling data set division results, each of which contains a first preset proportion of training data set and a second preset proportion of test data set.

[0088] Wherein, the first preset proportion and the second preset proportion can be set by those skilled in the art according to actual needs, and the present application embodiment does not make specific limitations. For example: the first preset proportion is set to 70%, and the second preset proportion is set to 30%; the first preset proportion is set to 80%, and the second preset proportion is set to 20%, etc.

[0089] In an optional embodiment, when the structured modeling data set is divided using multiple random seeds to obtain multiple modeling data set division results, multiple different stratified sampling strategies can be used to stratify sample the structured modeling data set according to the preset key variables to obtain multiple modeling data set division results. This optional method can ensure that the training data set and the test data set are consistent in important feature distribution, enhancing the representativeness and comprehensiveness of model training.

[0090] In the present application embodiment, to avoid the randomness and bias that a single random seed may bring, multiple random seeds are used to divide the data set multiple times and conduct training experiments, the performance of the model under different divisions is comprehensively analyzed, the stability and generalization performance of the model are evaluated by statistics of the mean and variance of multiple training results, and false judgments caused by single division are avoided. For example: in the actual implementation process, 10 different random seeds can be selected for testing, and the average value is selected as the result for final evaluation.

[0091] Step 103: Determine a plurality of preset machine learning models.

[0092] Wherein, the machine learning model includes: a random forest regression model, a gradient boosting tree model and a classification and regression tree model.

[0093] In the embodiments of the present application, in order to fully exploit the nonlinear relationship between variables and improve the accuracy of the model, three types of mainstream regression algorithms are selected: random forest regression model, gradient boosting tree model and classification and regression tree model. Through multiple rounds of training and error evaluation, the performance of each algorithm on different indicators is compared, and the optimal model is selected as the final target machine learning model.

[0094] Step 104: For each machine learning model, grid search is performed on the machine learning model parameters to obtain multiple model parameter combinations.

[0095] In each model, the model performance is further improved through automatic parameter search, and then the model is subjected to K-fold cross-validation.

[0096] In an optional embodiment, the way of performing grid search on the machine learning model parameters to obtain multiple model parameter combinations for each machine learning model can include the following sub-steps:

[0097] Sub-step 1: receiving a user's selection operation on parameters;

[0098] Sub-step 2: determining whether the number of parameters selected by the user exceeds a preset number;

[0099] Sub-step 3: if not, performing permutation and combination on the parameters selected by the user to obtain multiple model parameter combinations;

[0100] Sub-step 4: if yes, randomly extracting a preset number of parameters from the parameters selected by the user to perform permutation and combination to obtain multiple model parameter combinations.

[0101] This step is related to the parameter grid search. For each algorithm (i.e., the machine learning model described in the embodiments of the present application), different parameter options can be customized, and the parameter search space can be constructed through permutation and combination to finally determine the optimal model parameter combination. If too many parameters are set, a certain number of them are randomly extracted for trial using random search. In actual implementation, the maximum number set in the program is 50, and more than 50 permutation and combinations are randomly extracted for calculation.

[0102] Taking the random forest algorithm in the example as an example, the grid search method is used to combine and try the main parameters, and the set parameter range includes:

[0103] Number of decision trees: 100, 200 and 300 trees are used for model training to compare the prediction effect of the model under different forest sizes. More trees can improve the stability of the model, but also increase the calculation cost.

[0104] Max depth: Set 5, 10, 15 layers and unlimited depth four options, control model complexity. Shallow tree helps to prevent overfitting, deeper trees can more fully fit the training data.

[0105] Minimum split samples: Set to 2, 5 and 10, indicating the minimum number of samples required before each internal node is divided. Larger values will make the tree grow more conservative, helping to reduce the risk of overfitting.

[0106] Each parameter is combined and run, a total of 3x4x3=36 different parameter combinations, the optimal parameter is the number of decision trees in the forest 200, the maximum depth is not limited, and the minimum split sample number is 2.

[0107] Step 105: For each model parameter combination, use K-fold cross-validation method and multiple modeling dataset division results to train the machine learning model corresponding to the model parameter combination, and obtain the trained machine learning model; test the trained machine learning model through the test dataset, and obtain the evaluation index.

[0108] Use K-fold cross-validation method, K can be freely selected. For example, K=5 can be selected, the training set is divided into 5 equal size subsets, each round uses one of the subsets as the validation set, and the remaining 4 subsets as the training data, which is cycled 5 times to ensure that each data sample can be used as a validation set to participate in evaluation once.

[0109] In an optional embodiment, for each model parameter combination, the K-fold cross-validation method and multiple modeling dataset division results are used to train the machine learning model corresponding to the model parameter combination, and the trained machine learning model is obtained. The way can include the following sub-steps:

[0110] Determine the value of K in the K-fold cross-validation method; divide the training dataset into K equal size subsets; each round of training of the machine model corresponding to the model parameter combination uses one of the subsets as the test dataset, and the remaining K-1 subsets as the training dataset for single round training; cycle K times to train the machine model corresponding to the model parameter combination, and obtain the trained machine learning model.

[0111] The value of K can be flexibly set by those skilled in the art, and is not limited to the above-mentioned 5, but can also be 4 or 6, etc., which is not specifically limited in the embodiment of the present application.

[0112] Step 106: Determine the target machine learning model based on the evaluation index corresponding to each model parameter combination.

[0113] In an optional embodiment, the target machine learning model can be determined based on the evaluation index corresponding to each model parameter combination in the following manner: for each machine learning model, the optimal model parameter combination corresponding to the machine learning model is determined based on the evaluation index corresponding to each model parameter combination, and a first target machine learning model is generated; and the evaluation index optimal model is selected from the first target machine learning models as the target machine learning model according to the evaluation index.

[0114] After the grid search and cross-validation are completed to determine the target machine learning model, the joblib tool can be used to save the target machine learning model based on the optimal parameter configuration of each model, so that the target machine learning model can be used for subsequent calculation, online deployment or system integration. The joblib is a powerful Python library that provides simple and easy parallel computing and data persistence functions.

[0115] The evaluation index can include but is not limited to the determination coefficient, TRE (Total Relative Error), MSE (Mean System Error) and MPSE (Mean Percent Standard Error). When selecting the target machine learning model, the determination coefficient R 2 The first target machine learning model corresponding to the evaluation index with high determination coefficient, low TRE, low MSE and low MPSE is the optimal model, i.e., the target machine learning model, and the optimal model has strong robustness.

[0116] Step 107: training the target machine learning model based on the structured modeling data set to generate an accumulation amount estimation model.

[0117] The airborne lidar forest stock estimation method based on machine learning provided in this application generates a structured modeling dataset based on cloud data of various types of sample plots collected by airborne lidar and sample plot indicators obtained from actual surveys of each sample plot. The structured modeling dataset is then partitioned using multiple random seeds to obtain various modeling dataset partitioning results. Multiple preset machine learning models are determined. For each machine learning model, a grid search is performed on the machine learning model parameters to obtain multiple model parameter combinations. For each model parameter combination, the machine learning model corresponding to the model parameter combination is trained using K-fold cross-validation and the multiple modeling dataset partitioning results to obtain a trained machine learning model. The trained machine learning model is tested using a test dataset to obtain evaluation metrics. Based on the evaluation metrics corresponding to each model parameter combination, a target machine learning model is determined. The target machine learning model is trained based on the structured modeling dataset to generate a stock estimation model. The method provided in this application has the following advantages: First, it supports modeling without removing abnormal point cloud data, making full use of the original data and ensuring that the volume estimation model maintains good performance even when extreme samples are included, thereby improving the robustness and applicability of the volume estimation model's prediction results. Second, it eliminates the need for classification modeling of various types of sample plots, making it applicable to multiple forest types and exhibiting good universality. Third, by employing multiple random seeds and multi-parameter combinations for grid parameter search, the reliability and stability of the volume estimation model can be further enhanced.

[0118] The following is in conjunction with the appendix Figure 2 This paper uses a specific example to illustrate the airborne lidar forest stock estimation method based on machine learning provided in this application.

[0119] like Figure 2 The flowchart of the airborne lidar forest stock estimation method based on machine learning is shown in the figure. It mainly includes five parts: data acquisition, data preprocessing and variable extraction, sample system construction, model training and optimization, and model evaluation and application.

[0120] Data acquisition, data preprocessing, and variable extraction:

[0121] Based on airborne lidar point cloud data, this application first preprocesses the original point cloud using specialized software, including operations such as adjustment and normalization, and then extracts multiple structural feature variables to determine the explanatory variables.

[0122] These variables cover point cloud height and intensity indices, reflecting the vertical structure and reflectivity of forest stands, and providing rich high-dimensional data support for subsequent model building.

[0123] Constructing a sample system:

[0124] To improve the generality and stability of the model, a unified sample system is established in the present application, and the sample data of different forest types are integrated into a whole sample set. In this process, the sample plots are not pre-classified, and no outliers are excluded, avoiding the bias and information loss caused by classification, and ensuring the integrity and authenticity of the data. The sample data is randomly divided into training set and test set according to a certain proportion, ensuring the independence and representativeness of training and testing, thereby improving the generalization ability of the model.

[0125] Model training and optimization:

[0126] Based on the training data, the present application introduces various typical machine learning algorithms such as random forest, classification and regression tree and gradient boosting tree, optimizes the configuration of key parameters through grid search combined with cross-validation, improves the nonlinear expression ability and adaptability of the model, enhances the modeling ability of complex forest structure, prevents overfitting, and ensures the accuracy and stability of model prediction.

[0127] Model evaluation and application:

[0128] In order to comprehensively evaluate the performance of the model, a multi-dimensional error evaluation system is established, and the coefficient of determination (R 2 ), mean prediction error, systematic error and mean percentage standard error are used to evaluate the performance of the model on the test set. Finally, the model with the best performance is selected for practical application according to the indicators, ensuring that the model reaches an excellent level in accuracy, stability and interpretability.

[0129] The following will be combined with specific examples to explain the airborne laser radar forest volume estimation method based on machine learning of the embodiments of the present application.

[0130] In this specific example, 821 sample plot airborne laser radar point cloud data and measured forest volume data from the northeast forest area of China are used. In addition to the basic attribute information, 98 point cloud statistical variables are extracted from each sample plot, including 56 height variables and 42 intensity variables. Before modeling, the sample plot data is not pre-classified, and all sample plots are divided into training set and test set according to the ratio of 8:2. In the selection of regression algorithm, three models of random forest, CART decision tree and gradient boosting tree are constructed respectively, and the main hyperparameters are adjusted by grid search method, and the optimal parameter combination is determined by five-fold cross-validation to improve the stability and generalization ability of the model. The modeling results show that the three machine learning algorithms all achieve high fitting accuracy, among which the gradient boosting tree has the best effect, achieving R 2The average percentage standard error is as low as 8.10%, and the model variable importance ranking shows that X30, X31, X15, X33 and X14 are the five variables that have the greatest impact on the forest stock volume. Through examples, the effectiveness of the high-dimensional point cloud variable modeling approach adopted is verified. The trained stock volume estimation model is not only suitable for various forest types such as larch, Korean pine, birch and poplar, but also has good generalization ability and can be applied to forest resource survey and remote sensing interpretation practices in a wider range.

[0131] The machine learning-based airborne laser radar forest stock volume estimation method provided by the embodiments of the present application can significantly improve the accuracy of forest stock volume estimation, the model determination coefficient (R 2 ) can reach 0.972, and the average percentage standard error is as low as 8.10%. In the second aspect, the modeling method without rejecting abnormal points is supported, the original data is fully utilized, the model still maintains good performance in the case of containing extreme samples, and the robustness and application range of the results are improved. In the third aspect, the modeling without classification can be applied to various forest types, enhancing the universality and generalization ability of the model. In the fourth aspect, multiple random seeds and grid search with multiple parameter combinations are used to enhance the stability of model selection. In the fifth aspect, the modeling efficiency of high-dimensional variables of the laser radar is effectively improved, the key variables are automatically selected, and the manual intervention is reduced. In the sixth aspect, a unified and automated modeling process and parameter optimization mechanism are established to improve the modeling efficiency and model stability.

[0132] Figure 3 To achieve the machine learning-based airborne laser radar forest stock volume estimation device provided by the embodiments of the present application, a structural block diagram of the device is provided.

[0133] The machine learning-based airborne laser radar forest stock volume estimation device provided by the embodiments of the present application includes the following functional modules.

[0134] The generating module 301 is configured to generate a structured modeling data set according to the point cloud data of each type of sample site collected by the airborne laser radar and the sample site indicators obtained by the actual investigation of each sample site.

[0135] The dividing module 302 is configured to divide the structured modeling data set by using multiple random seeds to obtain multiple modeling data set division results, each of which contains a first preset proportion of training data sets and a second preset proportion of test data sets.

[0136] The determining module 303 is configured to determine a plurality of preset machine learning models, wherein the machine learning models include a random forest regression model, a gradient boosting tree model and a classification and regression tree model.

[0137] The search module 304 is configured to perform a grid search on the machine learning model parameters of each machine learning model to obtain a plurality of model parameter combinations.

[0138] The cross-validation module 305 is configured to train, for each model parameter combination, a machine learning model corresponding to the model parameter combination by using a K-fold cross-validation method and the plurality of modeling dataset division results to obtain a trained machine learning model, and test the trained machine learning model by using the test dataset to obtain an evaluation index.

[0139] The screening module 306 is configured to determine a target machine learning model based on the evaluation index corresponding to each model parameter combination.

[0140] The training module 307 is configured to train the target machine learning model based on the structured modeling dataset to generate an accumulation amount estimation model.

[0141] Optionally, the generation module comprises:

[0142] The first sub-module is configured to, for each plot, convert, after strip splicing and geometric correction of point cloud data collected by an airborne laser radar based on a flight control point and a ground control point, the point cloud data to a specific projection coordinate system.

[0143] The second sub-module is configured to correct laser intensity by using a flight height, a scanning angle, and a ground object reflectivity.

[0144] The third sub-module is configured to remove abnormal point cloud data according to a set height and reflectivity threshold to obtain preprocessed high-dimensional point cloud data.

[0145] The fourth sub-module is configured to extract a preset structural feature variable from the high-dimensional point cloud data to obtain high-precision remote sensing point cloud data.

[0146] Optionally, the division module is specifically configured to:

[0147] The structured modeling dataset is sampled in a plurality of different stratified sampling strategies according to a preset key variable to obtain a plurality of modeling dataset division results.

[0148] Optionally, the search module comprises:

[0149] The fifth sub-module is configured to receive a selection operation of a user on parameters.

[0150] The sixth sub-module is configured to determine whether a number of parameters selected by the user exceeds a preset number.

[0151] The seventh sub-module is configured to, if the number of parameters selected by the user does not exceed the preset number, arrange and combine the parameters selected by the user to obtain a plurality of model parameter combinations.

[0152] an eighth sub-module configured to, if the selected parameter is exceeded, randomly select a preset number of parameters from the selected parameters and arrange and combine the selected parameters to obtain a plurality of model parameter combinations.

[0153] Optionally, the cross-validation module is specifically configured to:

[0154] determine a value of K in the K-fold cross-validation method;

[0155] divide the training data set into K equal-sized subsets;

[0156] in each round of training of the machine learning model corresponding to the model parameter combination, use one of the subsets as a test data set and use the remaining K-1 subsets as training data sets for single-round training;

[0157] train the machine learning model corresponding to the model parameter combination for K times to obtain a trained machine learning model.

[0158] Optionally, the screening module is specifically configured to:

[0159] for each machine learning model, determine an optimal model parameter combination corresponding to the machine learning model based on the evaluation indexes corresponding to various model parameter combinations to generate a first target machine learning model;

[0160] select an evaluation index optimal model from the first target machine learning models as a target machine learning model according to the evaluation indexes.

[0161] The method embodiment provided by the device for estimating forest volume based on machine learning can achieve the processes of the method embodiment, and thus repeated descriptions are omitted here. Figure 3 The device for estimating forest volume based on machine learning can achieve the processes of the method embodiment, and thus repeated descriptions are omitted here. Figure 1 The method embodiment provided by the device for estimating forest volume based on machine learning can achieve the processes of the method embodiment, and thus repeated descriptions are omitted here.

[0162] The device for estimating forest volume based on machine learning provided by the embodiment of the present application has the following advantages.

[0163] Optionally, the embodiment of the present application further provides an electronic device, comprising a processor, a memory, a program or instructions stored in the memory and executable on the processor, the program or instructions being executed by the processor to implement the processes performed by the above-mentioned airborne laser radar forest volume estimation device based on machine learning, and achieve the same technical effects. To avoid repetition, details are not described here.

[0164] a memory for storing a computer program;

[0165] a processor for executing the program stored in the memory to implement the airborne laser radar forest volume estimation method based on machine learning performed by the processor shown in the above-mentioned method embodiment.

[0166] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is shown in the figure, but it does not mean that there is only one bus or only one type of bus.

[0167] The communication interface is used for communication between the above-mentioned terminal and other devices.

[0168] The memory can include a Random Access Memory (RAM) and can also include a non-volatile memory, such as at least one disk memory. Optionally, the memory can also be at least one storage device located away from the aforementioned processor.

[0169] The processor mentioned above can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; can also be a Digital Signal Processing (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

[0170] It is to be appreciated that the term "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0171] The above description is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered within the scope of protection of the present application.

Claims

1. A machine learning based airborne LiDAR forest volume estimation method, characterized in that, The method comprises: According to the point cloud data of various types of sample sites collected by the airborne laser radar and the sample site indexes obtained by the actual investigation of various sample sites, a structured modeling data set is generated; The structured modeling data set is divided by using multiple random seeds to obtain multiple modeling data set division results, and each modeling data set division result contains a first preset proportion of training data set and a second preset proportion of test data set; A plurality of preset machine learning models are determined, wherein the machine learning models include: a random forest regression model, a gradient boosting tree model, and a classification and regression tree model; For each machine learning model, grid search is performed on the machine learning model parameters to obtain multiple model parameter combinations; For each model parameter combination, the machine learning model corresponding to the model parameter combination is trained by using the K-fold cross-validation method and the multiple modeling data set division results to obtain a trained machine learning model; the trained machine learning model is tested by using the test data set to obtain an evaluation index; Based on the evaluation index corresponding to each model parameter combination, a target machine learning model is determined; Based on the structured modeling data set, the target machine learning model is trained to generate an accumulation amount estimation model.

2. The method of claim 1, wherein, The step of generating a structured modeling data set according to the point cloud data of various types of sample sites collected by the airborne laser radar and the sample site indexes obtained by the actual investigation of various sample sites comprises: For each sample site, after the point cloud data collected by the airborne laser radar is spliced and geometrically corrected based on the flight control points and the ground control points, it is converted to a specific projection coordinate system; The flight height, scanning angle and ground object reflectivity are used to correct the laser intensity; According to the set height and reflectivity threshold, abnormal point cloud data is removed to obtain preprocessed high-dimensional point cloud data; From the high-dimensional point cloud data, pre-set structural feature variables are extracted to obtain high-precision remote sensing point cloud data.

3. The method of claim 1, wherein, The step of dividing the structured modeling data set by using multiple random seeds to obtain multiple modeling data set division results comprises: According to the pre-set key variables, the structured modeling data set is stratified sampled by using multiple different stratified sampling strategies to obtain multiple modeling data set division results.

4. The method of claim 1, wherein, The step of performing grid search on the machine learning model parameters to obtain multiple model parameter combinations for each machine learning model comprises: Receiving a user's selection operation on parameters; Judging whether the number of parameters selected by the user exceeds a preset number; If not, the parameters selected by the user are arranged and combined to obtain multiple model parameter combinations; If so, a preset number of parameters selected by the user are randomly extracted for arrangement and combination to obtain multiple model parameter combinations.

5. The method of claim 1, wherein, The step of training the machine learning model corresponding to each model parameter combination by using the K-fold cross-validation method and the multiple modeling data set division results to obtain a trained machine learning model comprises: Determining the K value in the K-fold cross-validation method; Dividing the training data set into K equal-sized subsets; Each time a machine model corresponding to the model parameter combination is trained, one of the subsets is used as a test data set, and the remaining K-1 subsets are used as training data sets for single-round training; The machine learning model is trained K times to obtain a trained machine learning model.

6. The method of claim 1, wherein, The step of determining the target machine learning model based on the evaluation indexes corresponding to each model parameter combination includes: For each machine learning model, the optimal model parameter combination corresponding to the machine learning model is determined based on the evaluation indexes corresponding to each model parameter combination to generate a first target machine learning model; According to the evaluation indexes, an evaluation index optimal model is selected from the first target machine learning models as the target machine learning model.

7. An airborne LiDAR-based forest volume estimation apparatus based on machine learning, characterized by, The device includes: The generation module is configured to generate a structured modeling data set according to each type of sample point cloud data collected by the airborne laser radar and each sample site index obtained by actual sample site investigation; The division module is configured to divide the structured modeling data set using multiple random seeds to obtain multiple modeling data set division results, each of which contains a first preset proportion of training data sets and a second preset proportion of test data sets; The determination module is configured to determine a plurality of preset machine learning models, wherein the machine learning models include a random forest regression model, a gradient boosting tree model, and a classification and regression tree model; The search module is configured to perform grid search on the machine learning model parameters for each machine learning model to obtain multiple model parameter combinations; The cross-validation module is configured to train a machine learning model corresponding to each model parameter combination using a K-fold cross-validation method and the multiple modeling data set division results to obtain a trained machine learning model, and test the trained machine learning model using the test data set to obtain evaluation indexes; The screening module is configured to determine a target machine learning model based on the evaluation indexes corresponding to each model parameter combination; The training module is configured to train the target machine learning model based on the structured modeling data set to generate a storage capacity estimation model.

8. The apparatus of claim 7, wherein, The generation module includes: The first submodule is configured to convert the point cloud data collected by the airborne laser radar to a specific projection coordinate system after strip splicing and geometric correction based on the flight control points and the ground control points for each sample site; The second submodule is configured to correct the laser intensity using the flight height, the scanning angle, and the ground object reflectivity; The third submodule is configured to remove abnormal point cloud data according to the set height and reflectivity threshold to obtain preprocessed high-dimensional point cloud data; The fourth submodule is configured to extract a preset structural feature variable from the high-dimensional point cloud data to obtain high-precision remote sensing point cloud data.

9. The apparatus of claim 7, wherein, The search module includes: The fifth submodule is configured to receive a user selection operation on parameters; The sixth submodule is configured to determine whether the number of user-selected parameters exceeds a preset number; The seventh submodule is configured to arrange and combine the user-selected parameters to obtain multiple model parameter combinations if the number of user-selected parameters does not exceed the preset number. The eighth sub-module is configured to, if the selected parameters exceed the preset number, randomly select a preset number of parameters from the selected parameters and arrange and combine the selected parameters to obtain a plurality of combinations of model parameters.

10. An electronic device, comprising: The electronic device includes a processor, a memory, and a program or instructions stored on the memory and executable on the processor, and the processor executes the steps of any one of the forest volume estimation methods based on machine learning of claims 1-6.

Images